Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Yiyang Feng, Biddut Sarker Bijoy, Niranjan Balasubramanian, Jiawei Zhou
• Venue: arXiv: cs.AI
• Published: 2026-08-20
• Collected: 2026-08-21
• URL: http://arxiv.org/abs/2608.20274v1
• PDF: https://arxiv.org/pdf/2608.20274v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

Researchers studied how large language model (LLM) agents can learn and apply skills from completed tasks to perform better on future tasks. They found that inducing skills at the subtask level instead of at the task level, and using text-based skills instead of code-based ones, leads to more reliable skill transfer across tasks. The researchers also proposed a 'skill utility score' that can predict the success of skill transfer based on the specificity and abstractness of the induced skills.

What the Document Says

• Large language model (LLM) agents can learn from completed tasks to perform better on future tasks.
• Task-level skill induction mostly reduces agent's performance, while subtask-level skill induction raises it on average.
• Text skills transfer better than code skills.
• The researchers propose two properties for induced skills: specificity and abstractness. Neither property alone predicts task success, but their combined effect does.
• The proposed 'skill utility score' can predict the success of skill transfer based on skills and task descriptions without needing any task execution.

Key Concepts Mentioned

• large language model (LLM) agents
• skills induction
• task-level vs subtask-level
• text skills vs code skills
• skill utility score

Claims Made by the Document

• Induced skills can help LLM agents perform better with experience.
• Task-level skill induction reduces agent's performance below its no-memory baseline.
• Subtask-level skill induction raises the agent's performance above its no-memory baseline on average.
• Text skills transfer better than code skills.
• The combined effect of specificity and abstractness of induced skills predicts task success.

Evidence Used

• In practice, induced skills may transfer unreliably and can even harm the agent that retrieves them. When agent-induced skills transfer reliably across tasks remains an open question.
• To further understand our findings, we examine two complementary properties of the induced skills: specificity, which measures how closely a skill matches real tasks, and abstractness, which measures how evenly its relevance spreads across tasks.

Methods / Approach

• Comprehensive and controlled study comparing task-level with subtask-level skill induction and text with code skill formats

Limitations Stated by the Authors

The research focuses on understanding how the way skills are induced shapes their transfer across tasks, but it may not cover other factors that could affect skill transfer.

Limitations Noted by the Agent

Agent interpretation — not from the source.

As a text-based AI model, I don't have personal experiences or feelings. My responses are based on the data I've been trained on and my ability to process information provided to me.

Open Questions

• What other factors affect the reliability of skill transfer in LLM agents?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text doesn't specify whether the 'skill utility score' has been tested on various types of tasks or how it performs in real-world applications.

Raw Abstract / Excerpt

Large language model (LLM) agents can induce skills from completed tasks and reuse them later to grow more capable with experience. In practice, induced skills may transfer unreliably and can even harm the agent that retrieves them. When agent-induced skills transfer reliably across tasks remains an open question. We conduct a comprehensive and controlled study of how the way skills are induced shapes their transfer across tasks. Specifically, we compare task-level with subtask-level skill induction and text with code skill formats, the two axes along which existing methods differ. Task-level skills mostly reduce the agent's performance below its no-memory baseline while subtask-level skills raise it above on average, and text skills transfer better than code skills. To further understand our findings, we examine two complementary properties of the induced skills: specificity, which measures how closely a skill matches real tasks, and abstractness, which measures how evenly its relevance spreads across tasks. Neither property alone predicts task success, but their combined effect does, which we propose as a skill utility score. The score correlates consistently with task success when skills are transferred, and subtask-level and text skills score higher. Computing skill utility only needs the skills and task descriptions but not any task execution, so our score serves as a practical diagnostic of a skill memory before any new task runs.